Geographical analysis of cancer incidence in Ireland: A comparison of two Bayesian spatial models

Abstract 

Aims: Our objective was to describe the geographical variation in cancer incidence using gastro-intestinal and non-melanoma skin cancer incidence data in Ireland using two different Bayesian spatial models and to compare the performance of these models. Methods: Cases diagnosed between 1994 and 2003 were extracted from the National Cancer Registry of Ireland. Population data were estimated from census data. For each of 3401 electoral divisions (EDs), relative risk (RR) estimates were calculated and smoothed using a conditional autoregressive model (CAR) and a spatial partition model introduced by Hegarty and Barry using a product partition model (PPM). The results were compared by mapping the ratio of the two RR estimates and other mainly descriptive statistics. Results: The two methods gave broadly similar results. For gastro-intestinal cancers the RRs were lower in a northwest/southeast band across the country with greater RRs around Dublin, Cork and in Donegal. Greater RR of non-melanoma skin cancer was observed in coastal areas. Median differences between the RR estimates were small (=0.01). The range of RRs was wider when estimated by the CAR model illustrating that the PPM smoothed the data to a greater extent than the CAR model. Conclusions: The two approaches gave similar results providing stronger evidence for the resulting geographical patterns. PPMs give a more global picture of the risk distribution whereas CAR models provide more local estimates. The observed patterns may reflect socio-demographic or geographic variations in risk factors or access to cancer services. By helping to identify those risks, these maps may help in the optimal allocation of scarce health resources.

Keywords: Disease mapping, Bayesian spatial models, Conditional autoregressive model, Product partition model, Incidence, Gastro-intestinal cancer, Non-melanoma skin cancer